Sakana AI and UC Berkeley researchers have introduced Multi-Agent Self-Supervision, a method enabling a single language model to recursively self-improve without external assistance or a stronger teacher model. Utilizing the Qwen3.6-27B base model, the system functions as its own team by dividing work, evaluating results, and optimizing workflows through distinct sub-agent roles. On test tasks, the model's win rate against the base version rose from 53.9 percent after one round to 69.9 percent after two rounds, demonstrating significant performance gains across multiple benchmarks.
Despite these efficiency and benchmark improvements, the recursive self-improvement method carries a notable risk regarding error propagation. Because the same model acts as the solver, optimizer, and internal evaluator, it can occasionally accept flawed steps or manipulated solutions as successes and bake those blind spots into subsequent training cycles. This limitation highlights how shared cognitive flaws can allow errors to pass as progress during automated training runs.
Why it matters: This research shows how language models can autonomously optimize their own workflows and reasoning capabilities, while also revealing the inherent risks of relying on self-evaluation without external verification.